Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Single-step Controllable Music Bandwidth Extension With Flow Matching

Carlos Hernandez-Olivan, Hendrik Vincent Koops, Hao Hao Tan, Elio Quinton

TL;DR

The paper tackles the problem of restoring degraded music with controllable bandwidth extension by adapting FlowHigh to the music domain and introducing Dynamic Spectral Contour (DSC) as a time-varying control feature. It extends FlowHigh with audio-feature conditioning and classifier-free guidance to enable fine-grained, single-step restoration, reconstructing full-band audio from degraded inputs via a transformer-based vector-field estimator and a frozen BigVGAN vocoder. Empirical results show competitive performance against diffusion baselines in both full-band restoration and controlled restoration, with DSC providing the best balance between reconstruction quality and adherence to the target bandwidth. The work enables professionals to interactively steer restoration, offering a practical path toward more flexible and creative audio restoration workflows.

Abstract

Audio restoration consists in inverting degradations of a digital audio signal to recover what would have been the pristine quality signal before the degradation occurred. This is valuable in contexts such as archives of music recordings, particularly those of precious historical value, for which a clean version may have been lost or simply does not exist. Recent work applied generative models to audio restoration, showing promising improvement over previous methods, and opening the door to the ability to perform restoration operations that were not possible before. However, making these models finely controllable remains a challenge. In this paper, we propose an extension of FLowHigh and introduce the Dynamic Spectral Contour (DSC) as a control signal for bandwidth extension via classifier-free guidance. Our experiments show competitive model performance, and indicate that DSC is a promising feature to support fine-grained conditioning.

Single-step Controllable Music Bandwidth Extension With Flow Matching

TL;DR

The paper tackles the problem of restoring degraded music with controllable bandwidth extension by adapting FlowHigh to the music domain and introducing Dynamic Spectral Contour (DSC) as a time-varying control feature. It extends FlowHigh with audio-feature conditioning and classifier-free guidance to enable fine-grained, single-step restoration, reconstructing full-band audio from degraded inputs via a transformer-based vector-field estimator and a frozen BigVGAN vocoder. Empirical results show competitive performance against diffusion baselines in both full-band restoration and controlled restoration, with DSC providing the best balance between reconstruction quality and adherence to the target bandwidth. The work enables professionals to interactively steer restoration, offering a practical path toward more flexible and creative audio restoration workflows.

Abstract

Audio restoration consists in inverting degradations of a digital audio signal to recover what would have been the pristine quality signal before the degradation occurred. This is valuable in contexts such as archives of music recordings, particularly those of precious historical value, for which a clean version may have been lost or simply does not exist. Recent work applied generative models to audio restoration, showing promising improvement over previous methods, and opening the door to the ability to perform restoration operations that were not possible before. However, making these models finely controllable remains a challenge. In this paper, we propose an extension of FLowHigh and introduce the Dynamic Spectral Contour (DSC) as a control signal for bandwidth extension via classifier-free guidance. Our experiments show competitive model performance, and indicate that DSC is a promising feature to support fine-grained conditioning.
Paper Structure (14 sections, 1 equation, 3 figures, 3 tables)

This paper contains 14 sections, 1 equation, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Model Setup
  4. Controllability
  5. Audio Feature Control: Dynamic Spectral Contour
  6. Experiments and Results
  7. Dataset and preprocessing
  8. Evaluation
  9. Experiment 1: Full-band Restoration
  10. Results
  11. Experiment 2: Controlled Restoration
  12. Results
  13. Manipulating Dynamic Spectral Contour
  14. Conclusion

Figures (3)

  • Figure 1: Proposed Flow Matching method with audio controls for music restoration.
  • Figure 2: Comparison of bandwidth control features. Our proposed Dynamic Spectral Contour accurately captures bandwidth where spectral centroid and rolloff fail (e.g., in the low-energy region from 4s onwards).
  • Figure 3: Audio restoration effect of manipulating dsc control signals and a 4kHz degraded input. In purple, the dsc calculated from the original clean file. In green, dsc calculated from the restored audio using the manipulated dsc as control signal.